Klasifikasi Penyakit Paru-Paru dari Citra X-Ray Menggunakan EfficientNetB3 dengan Visualisasi Grad-CAM++

Abstract

Lung diseases, particularly Pneumonia, remain a major global health burden with high mortality rates, especially in developing countries. Chest X-ray interpretation is prone to subjectivity and radiologist errors, while visual overlap between Pneumonia and Lung opacity further complicates accurate diagnosis. This study aimed to develop an automated lung disease classification system using EfficientNet-B3 combined with Explainable Artificial Intelligence (XAI) to improve diagnostic transparency. The dataset comprised 2.952 chest X-ray images across three classes (Normal, Pneumonia, and Lung opacity). Preprocessing included resizing, normalization, augmentation, and two split scenarios: 70:15:15 and 80:10:10. The model was trained using transfer learning with pre-trained ImageNet weights over 50 epochs. Interpretability was assessed through Grad-CAM++ and LIME. Both configurations achieved strong performance with test accuracies of 92.57% and 92.93% and ROC-AUC values of 0.9875 and 0.9816, respectively. Lung opacity achieved the highest F1-score (>0.96), while Pneumonia recorded the lowest due to visual overlap. Grad-CAM++ and LIME consistently validated clinically relevant feature representations for Lung opacity, while inconsistencies in Normal and Pneumonia classes indicated residual dependence on non-specific features. EfficientNet-B3 demonstrated robust performance for lung disease classification, and the integration of Grad-CAM++ and LIME effectively enhanced model transparency to support radiological decision-making.